Python新手搭建LSTM自编码器遇模型未构建错误求助

我是Python新手，尝试用以下代码搭建LSTM自编码器，目标提取瓶颈层的压缩数据。但运行到编码器部分时，总会报错：This model has not yet been built. Build the model first by calling build()`，试过Stack Overflow上的一些建议，还是反复出错，请问哪里漏了？

完整代码 #

from google.colab import drive
drive.mount('/content/drive')
data=pd.read_csv(r'/content/drive/MyDrive/Datasets/All_Autoencoder_Data.csv')
data1 = data.drop('Labels', axis=1)
data1.shape
(225711, 68)

def create_sequences(X, time_steps=5):
    Xs = []
    for i in range(0, len(X)-time_steps, 5):
        Xs.append(X.iloc[i:(i+time_steps)].values)
    
    return np.array(Xs)

X_train=create_sequences(data1)
X_train.shape
(45142, 5, 68)

inputs = Input(shape=(X_train.shape[1], X_train.shape[2]))
eL0 = LSTM(68, activation='tanh', return_sequences=True,
            recurrent_activation="sigmoid", 
            kernel_initializer="glorot_uniform",
            recurrent_regularizer=regularizers.l2(0.001), 
            kernel_regularizer=regularizers.l2(0.001))(inputs)

eL1 = LSTM(32, activation='tanh', return_sequences=True,
            recurrent_activation="sigmoid",  
            kernel_initializer="glorot_uniform", 
            recurrent_regularizer=regularizers.l2(0.001), 
            kernel_regularizer=regularizers.l2(0.001))(eL0)

eL2 = LSTM(8, activation='tanh', return_sequences=False)(eL1)
#     eL3 = LSTM(16, activation='relu', return_sequences=False)(eL2)
h = RepeatVector(X_train.shape[1])(eL2)

dL2 = LSTM(32, activation='tanh', return_sequences=True)(h)

dL3 = LSTM(68, activation='tanh', return_sequences=True)(dL2)

output = TimeDistributed(Dense(X_train.shape[2]))(dL3)    
#output = Dense(X_train.shape[2])(dL4)  
model = Model(inputs=inputs, outputs=output)
#return model
model.summary()

model.build(input_shape=(None, 5, 68))

# fit the model to the data
callback = tf.keras.callbacks.EarlyStopping(monitor='loss', patience=3)
nb_epochs = 200
batch_size = 250
opt = keras.optimizers.Adam(learning_rate=0.001)
model.compile(optimizer=opt, loss='log_cosh', metrics=['accuracy'])

history = model.fit(X_train, X_train, 
                    epochs=nb_epochs, 
                    batch_size=batch_size,
                    validation_split=0.3, 
                    callbacks=[callback]
                    ).history



# plot the training losses
fig, ax = plt.subplots(figsize=(6, 4), dpi=80)
ax.plot(history['loss'], 'b', label='Train', linewidth=2)
ax.plot(history['val_loss'], 'r', label='Validation', linewidth=2)
ax.set_title('Model loss', fontsize=16)
ax.set_ylabel('Loss (mse)')
ax.set_xlabel('Epoch')
ax.legend(loc='upper right')
plt.show()

#Encoder
enc=Sequential()
enc.add(model.layers[1])
enc.add(model.layers[2])
enc.add(model.layers[3])
enc.summary()

报错信息 #

ValueError                                Traceback (most recent call last)
<ipython-input-50-225773931044> in <module>
      3 enc.add(model.layers[2])
      4 enc.add(model.layers[3])
----> 5 enc.summary()

/usr/local/lib/python3.9/dist-packages/keras/engine/training.py in summary(self, line_length, positions, print_fn, expand_nested, show_trainable, layer_range)
   3290         """
   3291         if not self.built:
-> 3292             raise ValueError(
   3293                 "This model has not yet been built. "
   3294                 "Build the model first by calling `build()` or by calling "

ValueError: This model has not yet been built. Build the model first by calling `build()` or by calling the model on a batch of data.

问题根源 #

你直接从训练好的自编码器中提取层到新的Sequential模型，但这些层的输入维度没有被新模型继承，导致Keras无法确定编码器的输入形状，因此模型未完成构建。

方法一：直接基于原模型创建编码器（推荐） #

不需要重新搭建Sequential，直接用Model类指定原模型的输入和瓶颈层输出，这样能完整继承原模型的层参数和输入形状：

# 瓶颈层是原模型的第4层（索引3），也就是eL2
encoder = Model(inputs=model.input, outputs=model.layers[3].output)
encoder.summary()

# 提取压缩数据
compressed_data = encoder.predict(X_train)

方法二：给Sequential编码器指定输入形状 #

如果一定要用Sequential，必须先明确输入形状，要么添加Input层，要么手动调用build方法：

enc = Sequential()
# 添加输入层，和原模型输入形状一致
enc.add(Input(shape=(X_train.shape[1], X_train.shape[2])))
enc.add(model.layers[1])
enc.add(model.layers[2])
enc.add(model.layers[3])
enc.summary()

# 或者不添加Input层，手动构建
# enc.build(input_shape=(None, X_train.shape[1], X_train.shape[2]))
# enc.summary()

内容的提问来源于stack exchange，提问作者Frenzy